What Do AIWP Models "Know" About Explosive Cyclogenesis?

AIWP (Artificial Intelligence Weather Prediction) has been rapidly advancing over the past several years, and now includes real-time models from Google DeepMind (GraphCast), NVIDIA (FourCastNet), Huawei ( Pangu Weather) and Microsoft (ClimaX), among others (see https://www.cira.colostate.edu/ml/ai4nwp/). These models have been trained on publicly available reanalysis data [ERA5, from the European Centre for Medium-Range Weather Forecasts (ECMWF)] and in terms of global aggregate scores like RMSE, are competitive with or in some cases performing better than NWP models from major centers such as ECMWF and NOAA’s Environmental Modeling Center (EMC).

Although AIWP models are not constrained by physics, but rather “learn” patterns from analysis data, some studies have suggested that these models can “learn” physics in the same way. We put this question to the test, in the context of explosive cyclogenesis events for the 2023-24 cold season in the Northern Hemispheric Atlantic basin. We perform diagnostics of AIWP model forecasts from one of the above models (GraphCast), initialized with the GEFS and run in real-time by the Cooperative Institute for Research in the Atmosphere (CIRA), for a selection of these events, and compare these results against analysis data.